AI in day-to-day engineering: where it helps and where I stay careful
How I use AI coding agents, generative AI and machine learning in automotive tool development, and the rules I follow for testing and data protection.
AI is part of how I build software today. I use coding agents every day, I have built generative-AI and machine-learning tools for engineering teams, and I hold ISTQB certifications in AI testing and testing with generative AI. This article summarizes what has worked for me, and where I deliberately stay careful.
1. Coding agents as a daily tool
I work with Claude Code, GitHub Copilot and Gemini CLI. They are most useful for:
- scaffolding new modules and tests from a clear description,
- refactoring and explaining unfamiliar legacy code,
- writing documentation and keeping it in sync with the code.
My rule: the agent proposes, I decide. Every change is reviewed and tested like a colleague's pull request. I also package recurring workflows as reusable Claude skills, so the same steps run the same way every time.
2. Generative AI for architecture checks
In the AI4ALL pipeline at Valeo, I built a tool that compares system-architecture elements from Polarion via its REST API and uses the Gemini API to flag inconsistencies and generate test cases. A Gradio front end makes it usable for engineers who never touch the code. The value is earlier feedback: inconsistencies surface during design instead of during integration testing.
3. Classic machine learning where it fits
Not every problem needs a large language model. For forecasting remaining engineering effort, a gradient-boosting regressor (scikit-learn) on historical CRETA exports, with a Streamlit dashboard, was the simpler and more explainable choice. Clean data pipelines and clear interfaces mattered more than the model itself.
4. Testing systems that are not deterministic
An AI component can give different answers to the same input, so classic "expected output equals actual output" tests are not enough. What I apply, in line with ISTQB CT-AI and CT-GenAI:
- fixed evaluation sets that represent real use, re-run after every change,
- checks on properties (correct facts, allowed format, no invented values) instead of exact wording,
- a human in the loop wherever a wrong answer has real consequences.
5. Data protection first
Automotive engineering data is confidential. Company data only goes to approved enterprise setups, such as Gemini through the company's Google Cloud project, never to personal accounts. The assistant on this website follows the same idea: it answers only from the public content of this site.
What I learned
- AI speeds up the work around the code (boilerplate, tests, documentation) more reliably than the hard design decisions.
- Choose the simplest model that works. Often that is not an LLM.
- Test behavior, not wording. That makes AI features testable and maintainable.